The bootstrap: a technique for data-driven statistics. Using computer-intensive analyses to explore experimental data

BACKGROUND

The concept of resampling data--more commonly referred to as bootstrapping--has been in use for more than three decades. Bootstrapping has considerable theoretical advantages when it is applied to non-Gaussian data. Most of the published literature is concerned with the mathematical aspects of the bootstrap but increasingly this technique is being utilized in medical and other fields.

METHODS

I reviewed the published literature following a 1994 publication assessing the transfer of technology, including the bootstrap, to the biomedical literature.

RESULTS

In the ten-year period following that 1994 paper there were 1679 published references to the technique in Medline. In that same time period the following citations were found in the four major medical journals-British Medical Journal (48), JAMA (51), Lancet (52) and the New England Journal of Medicine (45).

CONTENT

I introduce the basic theory of the bootstrap, the jackknife, and permutation tests. The bootstrap is used to estimate the accuracy of an estimator such as the standard error, a confidence interval, or the bias of an estimator. The technique may be useful for analysing smallish expensive-to-collect data sets where prior information is sparse, distributional assumptions are unclear, and where further data may be difficult to acquire. Some of the elementary uses of bootstrapping are illustrated by considering the calculation of confidence intervals such as for reference ranges or for experimental data findings, hypothesis testing such as comparing experimental findings, linear regression, and correlation when studying association and prediction of variables, non-linear regression such as used in immunoassay techniques, and ROC curve processing.

CONCLUSIONS

These techniques can supplement current nonparametric statistical methods and should be included, where appropriate, in the armamentarium of data processing methodologies.

Proficiency testing of creatine kinase and creatine kinase-2: the experience of the Ontario Laboratory Proficiency Testing Program

The Ontario Laboratory Proficiency Testing Program has regularly monitored the analytical performance of total creatine kinase (CK) (≈230 participants) and CK isoenzyme-2 (CK-MB) (≈160 participants) throughout the entire province. Consistently, a wide dispersion of results has been observed not only between different analyzer systems but also among identical analyzers. Accordingly, the results of the last three proficiency surveys for these analytes were examined statistically to establish both the extent of these variations and the range of values reported for the male upper reference ranges. The results of many of the analyzer systems were significantly different from each other, as were many of the reference ranges. This unsatisfactory situation may only be remedied by the use of reference materials as shown by others. The consequences of these findings also effect the reliability of epidemiological surveys such as the WHO MONICA Project (Circulation 1994;90:583–612), which monitors deaths due to heart disease and includes cardiac enzyme results in its criteria.

Biochemical markers of myocardial damage

OBJECTIVE

To assess various biochemical markers of myocardial damage.

METHODS AND RESULTS

Before routinely using any test as a biochemical marker of myocardial damage, the published evidence for its diagnostic utility must be critically assessed. Such assessment includes receiver operator curve (ROC) curve analyses, confidence interval estimates of claimed sensitivity and specificity values, and the effects of testing in serial and parallel modes. It is also necessary to establish the test's rule-in (high specificity) and rule-out (high sensitivity) decision thresholds that may vary with time after the onset of symptoms. The spectrum of ischemic heart disease includes acute (sudden death, non-Q- and Q-wave infarctions) and chronic (stable, unstable, and variant angina) conditions. Biochemical markers of myocardial damage are of most value in the diagnosis of acute ischemic heart disease, although increasingly some of these markers are being found to possess a prognostic value in chronic ischemic heart disease. The markers of enzymatic activity include aspartate aminotransferase, creatine kinase (together with isoenzymes and isoforms), and lactate dehydrogenase and isoenzymes. Creatine kinase isoenzyme-2 may also be measured immunologically, and this type of assay is in increasing use both because of its speed and because its blood levels rise earlier than the corresponding activities. The commercially available nonenzymatic markers are myoglobin and troponin T; troponin I is expected to become available in late 1995. While myoglobin is a nonspecific indicator of myocardial damage, its diagnostic value is due to its early appearance in blood. Troponin T is more cardiac specific, but the published data appears to suggest that the cardiac specificity of troponin I is superior. Troponin levels become abnormal at about the same time after the onset of symptoms as mass assays of creatine kinase isoenzyme-2; therefore, they are not useful as early markers of myocardial damage.

CONCLUSION

The availability of these nonenzymatic markers of myocardial damage must force a reassessment of the continued use of the enzymatic markers. Are they necessary, and if so, which ones should be retained?